Dual-energy X-ray mineral detection method based on dual-backbone YOLOv11

Abstract

This invention discloses a dual-backbone-based YOLO's The dual-energy X-ray ore detection method of 11 includes the following steps: acquiring dual-energy X-ray images of copper ore; constructing a copper ore detection model; inputting the dual-energy X-ray images of copper ore into the copper ore detection model for detection, thereby completing the copper ore sorting; the copper ore detection model adopts a dual-core architecture. YOLO's 11-model; dual-backbone YOLO's Model 11 in the original YOLO's In the 11 model, an auxiliary backbone network is introduced to compensate for accuracy loss, and modulated deformable convolution is used to replace traditional convolution; a multi-scale feature fusion module is designed to enhance feature transfer between the two backbone networks; through structured pruning and auxiliary backbone network design, this invention significantly reduces the floating-point operation of the model.

Description

Technical Field

[0001] This invention relates to the field of ore sorting technology, specifically to a dual-energy X-ray ore detection method based on dual-backbone YOLOv11. Background Technology

[0002] Copper, as a crucial basic raw material, holds strategic significance for national economic development. Although my country possesses abundant copper resources, its deposits are small-scale, with a high proportion of low-grade ore, poor beneficiation capabilities, and a large proportion of associated minerals. There is an urgent need to accelerate the research and development of efficient beneficiation technologies for low-grade and difficult-to-process ores to improve mining efficiency and sustainable development capabilities. With the rapid development of artificial intelligence technology, X-ray-based XRT intelligent beneficiation technology has been widely applied in domestic mines. Machine vision inspection technology, as a core component of intelligent beneficiation, directly impacts the efficiency of the beneficiation process.

[0003] Currently, most research focuses on the classification and segmentation of ore images. While these methods have improved accuracy, their low processing efficiency makes them unsuitable for the real-time requirements of practical industrial applications. With the development of object detection technology, modern object detection models can now meet the real-time needs of industrial applications. For example, the Faster OreFSDet model achieves efficient ore particle detection and a lightweight ore sorting architecture (LOSN) with a small number of samples, enabling efficient sorting of coal gangue. Although the above studies have shown significant effectiveness in sorting distinctive minerals such as coal and tungsten, existing methods are limited when dealing with complex minerals like copper ore with fine-grained, non-uniformly distributed particles, due to insufficient feature extraction and high false negative rates for small particles. Furthermore, existing models are not fully applicable to resource-constrained industrial beneficiation environments, and there is still room for improvement in terms of volume reduction and inference efficiency. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a dual-energy X-ray ore detection method based on dual-backbone YOLOv11, which aims to overcome the deficiencies of existing target detection algorithms in terms of model complexity, computational resource consumption, and processing of complex ores.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a dual-energy X-ray ore detection method based on dual-backbone YOLOv11, comprising the following steps: acquiring dual-energy X-ray images of copper ore; constructing a copper ore detection model; inputting the dual-energy X-ray images of copper ore into the copper ore detection model for detection, thereby completing copper ore sorting; the copper ore detection model adopts a dual-backbone YOLOv11 model; the dual-backbone YOLOv11 model introduces an auxiliary backbone network into the backbone network of the original YOLOv11 model, and uses modulated deformable convolution to replace traditional convolution; a multi-scale feature fusion module is designed to enhance feature transfer between the dual backbone networks;

[0006] The dual-backbone YOLOv11 model consists of a dual-backbone network, a multi-scale feature fusion module, an intermediate processing module, and a detection head module. The dual-backbone network includes a main branch and an auxiliary branch. The dual-energy X-ray image of the copper ore is input into the dual-backbone network for processing. A multi-scale feature fusion module is set between the main branch and the auxiliary branch of the dual-backbone network to transfer features between the main branch and the auxiliary branch. The output of the dual-backbone network is input to the intermediate processing module for processing, and the output of the intermediate processing module is input to the detection head module for processing to output the copper ore sorting results.

[0007] Furthermore, the main branch of the dual-backbone network consists of a first convolutional layer, a first modulated deformable convolution, a first C3k2 module, a second modulated deformable convolution, a second C3k2 module, a third modulated deformable convolution, a third C3k2 module, a fourth modulated deformable convolution, and a fourth C3k2 module connected in sequence; the auxiliary branch of the dual-backbone network consists of a fifth modulated deformable convolution, a fifth C3k2 module, a sixth modulated deformable convolution, a sixth C3k2 module, a seventh modulated deformable convolution, a seventh C3k2 module, a spatial pyramid pooling module, and a C2PSA module connected in sequence.

[0008] The multi-scale feature fusion module consists of a first channel splitting module, a second channel splitting module, a third channel splitting module, a first feature fusion module, a second feature fusion module, and a third feature fusion module.

[0009] Furthermore, the dual-energy X-ray image of copper ore is input into a dual-backbone network for processing. A multi-scale feature fusion module is set up between the main and auxiliary branches of the dual-backbone network to facilitate feature transfer between them. The specific process is as follows: the dual-energy X-ray image of copper ore is processed sequentially through a first convolutional layer, a first modulated deformable convolution, a first C3k2 module, a second modulated deformable convolution, a second C3k2 module, a third modulated deformable convolution, a third C3k2 module, a fourth modulated deformable convolution, and a fourth C3k2 module; the output of the first C3k2 module is input into a fifth modulated deformable convolution for processing; the output of the second C3k2 module is input into a first channel splitting module for processing; the output of the third C3k2 module is input into a second channel splitting module for processing; the output of the fourth C3k2 module is input into a third channel splitting module for processing; the outputs of the first, second, and third channel splitting modules, along with the output of the fifth modulated deformable convolution, are input into a first feature fusion module for processing; the output of the first feature fusion module is then combined with the output of the fifth modulated deformable convolution. The outputs of the deformable convolutions are fed into the fifth C3k2 module for processing; the output of the fifth C3k2 module is fed into the sixth modulated deformable convolution for processing; the outputs of the second and third channel splitting modules and the sixth modulated deformable convolution are fed into the second feature fusion module for processing; the outputs of the second feature fusion module and the sixth modulated deformable convolution are fed into the sixth C3k2 module for processing; the output of the sixth C3k2 module is fed into the seventh modulated deformable convolution for processing; the output of the seventh modulated deformable convolution and the third channel splitting module are fed into the third feature fusion module for processing; the outputs of the third feature fusion module and the seventh modulated deformable convolution are fed into the seventh C3k2 module for processing; the outputs of the seventh C3k2 module and the fourth C3k2 module are fed into the spatial pyramid pooling module for processing; the output of the spatial pyramid pooling module is fed into the C2PSA module for processing; and the outputs of the sixth, seventh, and C2PSA modules are used as the outputs of the dual backbone network.

[0010] Furthermore, the intermediate processing module includes an eighth C3k2 module, a ninth C3k2 module, a tenth C3k2 module, an eleventh C3k2 module, an eighth modulated deformable convolution, and a ninth modulated deformable convolution. The processing flow of the intermediate processing module is as follows: the output of the C2PSA module is upsampled and concatenated with the seventh C3k2 module to obtain the first concatenated feature; the first concatenated feature is input into the eighth C3k2 module for processing; the output of the eighth C3k2 module is upsampled and concatenated with the output of the sixth C3k2 module to obtain the second concatenated feature; the second concatenated feature is then processed sequentially through the ninth C3k2 module. The C3k2 module and the eighth modulated deformable convolution are processed; the output of the eighth modulated deformable convolution and the output of the eighth C3k2 module are concatenated to obtain the third concatenated feature; the third concatenated feature is then processed sequentially through the tenth C3k2 module and the ninth modulated deformable convolution; the output of the ninth modulated deformable convolution and the output of the C2PSA module are concatenated to obtain the fourth concatenated feature; the fourth concatenated feature is then input into the eleventh C3k2 module for processing; the outputs of the ninth, tenth, and eleventh C3k2 modules are used as the outputs of the intermediate processing module.

[0011] Furthermore, the first, second, third, fourth, fifth, sixth, seventh, eighth, and ninth modulated deformable convolutions have the same structure; the processing procedure of the first modulated deformable convolution is expressed as follows:

[0012] ;

[0013] In the formula, Indicates the location within a regular grid. The output after the first modulated deformable convolution calculation; This represents the input to the first modulated deformable convolution, i.e., the output of the first convolutional layer; Indicates the position of the corresponding regular grid. The kernel parameters at the location; Indicates the position of the regular grid Corresponding position offset; Indicates a mask; This represents the set of regular grid locations.

[0014] Furthermore, the detection head module consists of multiple lightweight detection heads based on a parameter sharing strategy; each lightweight detection head based on the parameter sharing strategy receives the outputs of the ninth C3k2 module, the tenth C3k2 module, and the eleventh C3k2 module respectively for copper ore sorting.

[0015] The lightweight detection head based on the parameter sharing strategy includes a second depthwise separable convolution, a fifth convolutional layer, a first two-dimensional convolutional layer, a third depthwise separable convolution, a sixth convolutional layer, and a second two-dimensional convolutional layer. The processing flow of the lightweight detection head based on the parameter sharing strategy is as follows: the input is initially processed by the second depthwise separable convolution and the fifth convolutional layer, where the output of the fifth convolutional layer is sequentially fed into the regression branch and the classification branch; the regression branch generates the regression prediction result of copper ore through the first two-dimensional convolutional layer; the classification branch passes the output of the fifth convolutional layer sequentially through the third depthwise separable convolution, the sixth convolutional layer, and the second two-dimensional convolutional layer to generate the classification prediction result of copper ore.

[0016] Furthermore, the first channel splitting module, the second channel splitting module, and the third channel splitting module have the same structure, each consisting of a first depthwise separable convolutional layer, a fourth convolutional layer, and a third segmentation layer connected in sequence.

[0017] The first feature fusion module, the second feature fusion module, and the third feature fusion module have the same structure; the processing flow is as follows: the input is summed after undergoing bilinear interpolation to obtain the output.

[0018] Furthermore, the first, second, fifth, eighth, ninth, and tenth C3k2 modules have the same structure. The first C3k2 module includes a first splicing layer, a first segmentation layer, a first bottleneck layer, a second bottleneck layer, a second splicing layer, and a second convolutional layer. The processing flow of the first C3k2 module is as follows: the output of the first modulated deformable convolution is processed sequentially through the first splicing layer, the first segmentation layer, the first bottleneck layer, and the second bottleneck layer; the output of the second bottleneck layer, together with the output of the first splicing layer and the output of the first segmentation layer, is input into the second splicing layer for splicing and then enters the second convolutional layer for processing to obtain the output of the first C3k2 module.

[0019] Furthermore, the third, fourth, sixth, seventh, and eleventh C3k2 modules have the same structure. The third C3k2 module includes a third splicing layer, a second segmentation layer, a first C3k module, a second C3k module, a fourth splicing layer, and a third convolutional layer. The processing flow of the third C3k2 module is as follows: the output of the first modulated deformable convolution is processed sequentially through the third splicing layer, the second segmentation layer, the first C3k module, and the second C3k module; the output of the second C3k module, the output of the third splicing layer, and the output of the second segmentation layer are input together into the fourth splicing layer and then processed in the third convolutional layer to obtain the output of the third C3k2 module.

[0020] Furthermore, the dual-backbone YOLOv11 model employs the SIoU loss function.

[0021] Compared with existing technologies, this invention has the following advantages: Addressing the computational overhead of the YOLOv11 model, this invention reduces computational burden, lowers the number of model parameters and computational load, improves inference speed, and reduces hardware resource consumption by optimizing the network structure through network pruning; by introducing an auxiliary backbone network to compensate for accuracy loss, it maintains detection accuracy while reducing computational load; by using modulated deformable convolution instead of traditional convolution, it improves the model's adaptability to ore morphology, enabling the model to better fit complex ore boundaries and irregular contour features; by designing a multi-scale feature fusion module to enhance feature transfer between the two backbone networks, it improves the detection effect of targets at different scales; by utilizing a parameter sharing strategy to lightweight the detection head, it reduces redundant parameters in the detection head, improving operational efficiency while maintaining accuracy; and by introducing the SIoU loss function, it makes target localization more accurate. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the copper ore detection model of the present invention.

[0023] Figure 2 This is a schematic diagram of the specific structure of the intermediate processing module and the detection head module of the present invention.

[0024] Figure 3 This is a schematic diagram of the lightweight detection head based on the parameter sharing strategy of the present invention. Detailed Implementation

[0025] This invention provides a technical solution: a dual-energy X-ray mineral detection method based on dual-backbone YOLOv11, comprising the following steps:

[0026] Acquire dual-energy X-ray images of copper ore; construct a copper ore detection model, such as... Figure 1 As shown, the dual-energy X-ray image of copper ore is input into the copper ore detection model for detection, and the copper ore sorting is completed. The copper ore detection model adopts the dual-backbone YOLOv11 model. The dual-backbone YOLOv11 model introduces an auxiliary backbone network to compensate for the accuracy loss in the backbone network of the original YOLOv11 model, and uses modulated deformable convolution to replace the traditional convolution. A multi-scale feature fusion module is designed to enhance the feature transfer between the dual backbone networks.

[0027] The dual-backbone YOLOv11 model consists of a two-backbone network, a multi-scale feature fusion module (MSFF), an intermediate processing module (Neck), and a detection head module. The two-backbone network includes a main branch and an auxiliary branch. The dual-energy X-ray image of the copper ore is input into the two-backbone network for processing. The multi-scale feature fusion module is set between the main branch and the auxiliary branch of the two-backbone network to transfer features between the main branch and the auxiliary branch. The output of the two-backbone network is input to the intermediate processing module for processing, and the output of the intermediate processing module is input to the detection head module for processing to output the copper ore sorting results.

[0028] The main branch of the dual-backbone network consists of a first convolutional layer, a first modulated deformable convolution (MDConv), a first C3k2 module, a second modulated deformable convolution, a second C3k2 module, a third modulated deformable convolution, a third C3k2 module, a fourth modulated deformable convolution, and a fourth C3k2 module connected in sequence. The auxiliary branch of the dual-backbone network consists of a fifth modulated deformable convolution, a fifth C3k2 module, a sixth modulated deformable convolution, a sixth C3k2 module, a seventh modulated deformable convolution, a seventh C3k2 module, a spatial pyramid pooling module (SPPF), and a C2PSA module connected in sequence.

[0029] The multi-scale feature fusion module consists of a first channel splitting module (Csplitter), a second channel splitting module, a third channel splitting module, a first feature fusion module, a second feature fusion module, and a third feature fusion module.

[0030] The process involves inputting dual-energy X-ray images of copper ore into a dual-backbone network for processing. A multi-scale feature fusion module is set up between the main and auxiliary branches of the dual-backbone network to facilitate feature transfer between them. The specific process is as follows: The dual-energy X-ray image of copper ore is sequentially processed through a first convolutional layer, a first modulated deformable convolution, a first C3k2 module, a second modulated deformable convolution, a second C3k2 module, a third modulated deformable convolution, a third C3k2 module, a fourth modulated deformable convolution, and a fourth C3k2 module. The output of the first C3k2 module is input into a fifth modulated deformable convolution for processing. The output of the second C3k2 module is input into a first channel splitting module for processing. The output of the third C3k2 module is input into a second channel splitting module for processing. The output of the fourth C3k2 module is input into a third channel splitting module for processing. The outputs of the first, second, and third channel splitting modules, along with the output of the fifth modulated deformable convolution, are input into a first feature fusion module for processing. The output of the first feature fusion module is then combined with the output of the fifth modulated deformable convolution. The outputs of the deformable convolutions are fed into the fifth C3k2 module for processing; the output of the fifth C3k2 module is fed into the sixth modulated deformable convolution for processing; the outputs of the second and third channel splitting modules and the sixth modulated deformable convolution are fed into the second feature fusion module for processing; the outputs of the second feature fusion module and the sixth modulated deformable convolution are fed into the sixth C3k2 module for processing; the output of the sixth C3k2 module is fed into the seventh modulated deformable convolution for processing; the output of the seventh modulated deformable convolution and the third channel splitting module are fed into the third feature fusion module for processing; the outputs of the third feature fusion module and the seventh modulated deformable convolution are fed into the seventh C3k2 module for processing; the outputs of the seventh C3k2 module and the fourth C3k2 module are fed into the spatial pyramid pooling module for processing; the output of the spatial pyramid pooling module is fed into the C2PSA module for processing; the outputs of the sixth, seventh, and C2PSA modules are used as the outputs of the dual backbone network.

[0031] like Figure 2As shown, the intermediate processing module includes an eighth C3k2 module, a ninth C3k2 module, a tenth C3k2 module, an eleventh C3k2 module, an eighth modulated deformable convolution, and a ninth modulated deformable convolution. The processing flow of the intermediate processing module is as follows: the output of the C2PSA module is upsampled and concatenated with the seventh C3k2 module to obtain the first concatenated feature; the first concatenated feature is input into the eighth C3k2 module for processing; the output of the eighth C3k2 module is upsampled and concatenated with the output of the sixth C3k2 module to obtain the second concatenated feature; the second concatenated feature is then processed sequentially through the ninth C3k2 module. The 3k2 module and the eighth modulated deformable convolution are processed; the output of the eighth modulated deformable convolution and the output of the eighth C3k2 module are concatenated to obtain the third concatenated feature; the third concatenated feature is then processed sequentially through the tenth C3k2 module and the ninth modulated deformable convolution; the output of the ninth modulated deformable convolution and the output of the C2PSA module are concatenated to obtain the fourth concatenated feature; the fourth concatenated feature is then input into the eleventh C3k2 module for processing; the outputs of the ninth, tenth, and eleventh C3k2 modules are used as the outputs of the intermediate processing module.

[0032] The detection head module consists of multiple lightweight detection heads (LDetect) based on a parameter sharing strategy. Each lightweight detection head (LDetect) based on the parameter sharing strategy receives the outputs of the ninth C3k2 module, the tenth C3k2 module, and the eleventh C3k2 module respectively for copper ore sorting.

[0033] The first, second, third, fourth, fifth, sixth, seventh, eighth, and ninth modulated deformable convolutions share the same structure. Modulated deformable convolutions achieve flexible sampling by introducing offset and mask mechanisms. In a modulated deformable convolution, each position has a corresponding offset, ensuring that the sampling position depends not only on its position within the regular grid but also on the offset, thus achieving irregular sampling. Simultaneously, the mask is a learned scalar used to adjust the weights of each sampling point. The processing of the first modulated deformable convolution can be represented as follows:

[0034] ;

[0035] In the formula, Indicates the location within a regular grid. The output after the first modulated deformable convolution calculation; This represents the input to the first modulated deformable convolution, i.e., the output of the first convolutional layer; Indicates the position of the corresponding regular grid. The kernel parameters at the location; Indicates the position of the regular grid Corresponding position offset; Indicates a mask; This represents the set of regular grid locations.

[0036] Since the second, third, fourth, fifth, sixth, seventh, eighth, and ninth modulated deformable convolutions have the same structure and processing flow as the first modulated deformable convolution, they will not be elaborated upon here.

[0037] The first, second, fifth, eighth, ninth, and tenth C3k2 modules have the same structure. The first C3k2 module includes a first splicing layer, a first splitting layer, a first bottleneck layer, a second bottleneck layer, a second splicing layer, and a second convolutional layer. The processing flow of the first C3k2 module is as follows: the output of the first modulated deformable convolution is processed sequentially through the first splicing layer, the first splitting layer, the first bottleneck layer, and the second bottleneck layer; the output of the second bottleneck layer, the output of the first splicing layer, and the output of the first splitting layer are input together into the second splicing layer and then processed into the second convolutional layer to obtain the output of the first C3k2 module.

[0038] The second, fifth, eighth, ninth, and tenth C3k2 modules have the same structure and processing flow as the first C3k2 module, so they will not be described in detail here.

[0039] The third, fourth, sixth, seventh, and eleventh C3k2 modules have the same structure. The third C3k2 module includes a third splicing layer, a second segmentation layer, a first c3k module (the C3k module is a standard module included in the YOLOv11 model and is existing technology), a second c3k module, a fourth splicing layer, and a third convolutional layer. The processing flow of the third C3k2 module is as follows: the output of the first modulated deformable convolution is processed sequentially through the third splicing layer, the second segmentation layer, the first c3k module, and the second c3k module; the output of the second c3k module, the output of the third splicing layer, and the output of the second segmentation layer are input together into the fourth splicing layer and then processed in the third convolutional layer to obtain the output of the third C3k2 module.

[0040] The fourth, sixth, seventh, and eleventh C3k2 modules have the same structure and processing flow as the third C3k2 module, so they will not be described in detail here.

[0041] The first channel splitting module (Csplitter), the second channel splitting module, and the third channel splitting module have the same structure, each consisting of a first depthwise separable convolution (DWConv), a fourth convolutional layer, and a third splitting layer connected in sequence.

[0042] The first feature fusion module (FFusion), the second feature fusion module and the third feature fusion module have the same structure; their processing flow is as follows: the input is summed after passing through bilinear interpolation to obtain the output; the first feature fusion module, the second feature fusion module and the third feature fusion module adopt the feature fusion method commonly used in this field, which belongs to the existing technology scope and is usually also called weighted feature fusion.

[0043] In the original YOLOv11 model's detection head, the original architecture used two independent convolutional branches to perform regression (bounding box prediction) and classification (object category prediction) tasks respectively. However, in ore detection tasks, X-ray transmission imaging of ore images typically has simple backgrounds, clear contrast between the ore and the background, and relatively clear features. Therefore, the complex convolutional structure did not effectively improve feature extraction but instead increased the computational burden. To address this, a lightweight detection head (LDetect) based on a parameter-sharing strategy was designed using a shared convolutional strategy as the detection head for the dual-backbone YOLOv11 model.

[0044] like Figure 3As shown, the lightweight detection head (LDetect) based on a parameter-sharing strategy includes a second depthwise separable convolution, a fifth convolutional layer, a first two-dimensional convolutional layer (Conv2d), a third depthwise separable convolution, a sixth convolutional layer, and a second two-dimensional convolutional layer. The processing flow of the lightweight detection head based on the parameter-sharing strategy is as follows: The input (the output of the ninth, tenth, or eleventh C3k2 module) first passes through the second depthwise separable convolution and the fifth convolutional layer for preliminary processing. The output of the fifth convolutional layer then enters the regression branch and the classification branch. The regression branch generates a regression prediction result for copper ore through the first two-dimensional convolutional layer, representing the location information of the target bounding box, including parameters such as center coordinates, width, and height. The classification branch passes the output of the fifth convolutional layer through the third depthwise separable convolution, the sixth convolutional layer, and the second two-dimensional convolutional layer to generate a classification prediction result for copper ore, representing the probability distribution of the category to which each candidate box belongs. This output is used to calculate the cross-entropy loss to optimize the model's classification accuracy. This detection head achieves joint optimization of target location and category through the synergistic effect of regression and classification branches, thereby improving overall detection performance. The regression prediction results correspond to the SIoU loss function, and the classification prediction results correspond to the cross-entropy loss function, thus clarifying the correspondence between each output and the loss function.

[0045] The original YOLOv11 model uses the CIoU loss function, which optimizes the overlap between predicted and ground truth bounding boxes based on the traditional IoU loss function, while also considering center point distance, aspect ratio consistency, and geometric shape matching. However, in ore detection tasks, due to the simple background and high contrast between the ore and the background, the CIoU loss function penalizes small deviations in highly overlapping boxes weakly, limiting the improvement of localization accuracy. Especially when the density difference within the ore is small or the edges are blurred, the evaluation capability of the CIoU loss function decreases, leading to unstable bounding box regression and affecting localization accuracy. To address the localization error problem in the bounding box regression process, the SioU loss function is introduced as the loss function for the dual-backbone YOLOv11 model. The SioU loss function comprehensively improves the accuracy and stability of the bounding box regression process by introducing angle loss, distance loss, shape loss, and IoU loss.

[0046] To further verify the effectiveness of the proposed dual-backbone YOLOv11 model in ore detection tasks, experiments were conducted using classic target detection models such as YOLOv5n, YOLOv8n, YOLOv10n, and YOLOv11n. The performance of each model was analyzed in detail using five metrics: mAP@0.5:0.95, FLOPs, Params, FPS, and Model Size. The results are shown in Table 1. Experimental results show that the proposed dual-backbone YOLOv11 model achieves an mAP@0.5:0.95 score of 0.977, outperforming other models and indicating its superior performance in both positioning accuracy and recognition capability. Furthermore, the proposed dual-backbone YOLOv11 model exhibits significant lightweight advantages in FLOPs, Params, and Model Size metrics, with lower computational and storage overhead, making it suitable for resource-constrained applications. Regarding detection speed, the proposed dual-backbone YOLOv11 model achieves an FPS of 56.7, demonstrating excellent real-time detection capabilities. In summary, the dual-backbone YOLOv11 model of this invention exhibits superior performance in terms of accuracy, efficiency, and resource consumption, making it suitable for ore detection tasks.

[0047] Table 1 Comparison results of each model

[0048]

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dual-energy X-ray mineral detection method based on dual-backbone YOLOv11, characterized in that, The process includes the following steps: acquiring dual-energy X-ray images of copper ore; constructing a copper ore detection model; inputting the dual-energy X-ray images of copper ore into the copper ore detection model for detection, thereby completing the copper ore sorting; the copper ore detection model adopts a dual-backbone YOLOv11 model; the dual-backbone YOLOv11 model introduces an auxiliary backbone network into the backbone network of the original YOLOv11 model, and uses modulated deformable convolution to replace the traditional convolution; a multi-scale feature fusion module is designed to enhance feature transfer between the dual backbone networks; The dual-backbone YOLOv11 model consists of a dual-backbone network, a multi-scale feature fusion module, an intermediate processing module, and a detection head module. The dual-backbone network includes a main branch and an auxiliary branch. The dual-energy X-ray image of the copper ore is input into the dual-backbone network for processing. A multi-scale feature fusion module is set between the main branch and the auxiliary branch of the dual-backbone network to transfer features between the main branch and the auxiliary branch. The output of the dual-backbone network is input to the intermediate processing module for processing, and the output of the intermediate processing module is input to the detection head module for processing to output the copper ore sorting results. The detection head module consists of multiple lightweight detection heads based on a parameter sharing strategy; the lightweight detection heads based on the parameter sharing strategy receive the output of the intermediate processing module to perform copper ore sorting; The lightweight detection head based on the parameter sharing strategy includes a second depthwise separable convolution, a fifth convolutional layer, a first two-dimensional convolutional layer, a third depthwise separable convolution, a sixth convolutional layer, and a second two-dimensional convolutional layer. The processing flow of the lightweight detection head based on the parameter sharing strategy is as follows: the input is initially processed by the second depthwise separable convolution and the fifth convolutional layer, where the output of the fifth convolutional layer is sequentially fed into the regression branch and the classification branch; the regression branch generates the regression prediction result of copper ore through the first two-dimensional convolutional layer; the classification branch passes the output of the fifth convolutional layer sequentially through the third depthwise separable convolution, the sixth convolutional layer, and the second two-dimensional convolutional layer to generate the classification prediction result of copper ore.

2. The dual-energy X-ray mineral detection method based on dual-backbone YOLOv11 according to claim 1, characterized in that: The main branch of the dual-backbone network consists of a first convolutional layer, a first modulated deformable convolution, a first C3k2 module, a second modulated deformable convolution, a second C3k2 module, a third modulated deformable convolution, a third C3k2 module, a fourth modulated deformable convolution, and a fourth C3k2 module, connected in sequence; the auxiliary branch of the dual-backbone network consists of a fifth modulated deformable convolution, a fifth C3k2 module, a sixth modulated deformable convolution, a sixth C3k2 module, a seventh modulated deformable convolution, a seventh C3k2 module, a spatial pyramid pooling module, and a C2PSA module, connected in sequence. The multi-scale feature fusion module consists of a first channel splitting module, a second channel splitting module, a third channel splitting module, a first feature fusion module, a second feature fusion module, and a third feature fusion module.

3. The dual-energy X-ray mineral detection method based on dual-backbone YOLOv11 according to claim 2, characterized in that: The dual-energy X-ray image of copper ore is input into a dual-backbone network for processing. A multi-scale feature fusion module is set up between the main and auxiliary branches of the dual-backbone network to facilitate feature transfer between them. The specific process is as follows: the dual-energy X-ray image of copper ore is sequentially processed through the first convolutional layer, the first modulated deformable convolution, the first C3k2 module, the second modulated deformable convolution, the second C3k2 module, the third modulated deformable convolution, the third C3k2 module, the fourth modulated deformable convolution, and the fourth C3k2 module; the output of the first C3k2 module is then input into the fifth modulated deformable convolution for further processing. The output of the second C3k2 module is input into the first channel splitting module for processing; The output of the third C3k2 module is input into the second channel splitting module for processing; The output of the fourth C3k2 module is input to the third channel split module for processing; The outputs of the first channel splitting module, the second channel splitting module, and the third channel splitting module, together with the output of the fifth modulated deformable convolution, are input into the first feature fusion module for processing. The output of the first feature fusion module, together with the output of the fifth modulated deformable convolution, is input into the fifth C3k2 module for processing. The output of the fifth C3k2 module is input into the sixth modulated deformable convolution for processing; the outputs of the second channel splitting module, the third channel splitting module, and the sixth modulated deformable convolution are input together into the second feature fusion module for processing; The output of the second feature fusion module and the output of the sixth modulated deformable convolution are input together into the sixth C3k2 module for processing; The output of the sixth C3k2 module is input into the seventh modulated deformable convolution for processing; the output of the seventh modulated deformable convolution and the output of the third channel splitting module are input into the third feature fusion module for processing. The output of the third feature fusion module and the output of the seventh modulated deformable convolution are input together into the seventh C3k2 module for processing; The outputs of the seventh C3k2 module and the fourth C3k2 module are input together into the spatial pyramid pooling module for processing; The output of the spatial pyramid pooling module is input into the C2PSA module for processing; The outputs of the sixth C3k2 module, the seventh C3k2 module, and the C2PSA module are used as the outputs of the dual backbone network.

4. The dual-energy X-ray mineral detection method based on dual-backbone YOLOv11 according to claim 3, characterized in that: The intermediate processing module includes the eighth C3k2 module, the ninth C3k2 module, the tenth C3k2 module, the eleventh C3k2 module, the eighth modulated deformable convolution, and the ninth modulated deformable convolution. The processing flow of the intermediate processing module is as follows: the output of the C2PSA module is upsampled and concatenated with the seventh C3k2 module to obtain the first concatenated feature; the first concatenated feature is input into the eighth C3k2 module for processing. The output of the eighth C3k2 module is upsampled and concatenated with the output of the sixth C3k2 module to obtain the second concatenated feature. The second concatenated feature is then processed sequentially through the ninth C3k2 module and the eighth modulated deformable convolution. The output of the eighth modulated deformable convolution and the output of the eighth C3k2 module are concatenated to obtain the third concatenated feature. The third concatenated feature is then processed sequentially through the tenth C3k2 module and the ninth modulated deformable convolution. The output of the ninth modulated deformable convolution and the output of the C2PSA module are then concatenated to obtain the fourth concatenated feature. The fourth concatenated feature is then input into the eleventh C3k2 module for processing. The outputs of the ninth, tenth, and eleventh C3k2 modules are used as the outputs of the intermediate processing module.

5. The dual-energy X-ray mineral detection method based on dual-backbone YOLOv11 according to claim 4, characterized in that: The first, second, third, fourth, fifth, sixth, seventh, eighth, and ninth modulated deformable convolutions have the same structure; the processing procedure of the first modulated deformable convolution is expressed as follows: ； In the formula, Indicates the location within the regular grid. The output after the first modulated deformable convolution calculation; This represents the input to the first modulated deformable convolution, i.e., the output of the first convolutional layer; Indicates the position of the corresponding regular grid. The kernel parameters at the location; Indicates the position of the regular grid Corresponding position offset; Indicates a mask; This represents the set of regular grid locations.

6. The dual-energy X-ray mineral detection method based on dual-backbone YOLOv11 according to claim 5, characterized in that: Each lightweight detection head based on the parameter sharing strategy receives the outputs of the ninth, tenth, and eleventh C3k2 modules to sort copper ore.

7. The dual-energy X-ray mineral detection method based on dual-backbone YOLOv11 according to claim 6, characterized in that: The first channel splitting module, the second channel splitting module, and the third channel splitting module have the same structure, each consisting of a first depthwise separable convolutional layer, a fourth convolutional layer, and a third segmentation layer connected in sequence. The first feature fusion module, the second feature fusion module, and the third feature fusion module have the same structure; the processing flow is as follows: the input is summed after undergoing bilinear interpolation to obtain the output.

8. The dual-energy X-ray mineral detection method based on dual-backbone YOLOv11 according to claim 7, characterized in that: The first, second, fifth, eighth, ninth, and tenth C3k2 modules have the same structure. The first C3k2 module includes a first splicing layer, a first segmentation layer, a first bottleneck layer, a second bottleneck layer, a second splicing layer, and a second convolutional layer. The processing flow of the first C3k2 module is as follows: the output of the first modulated deformable convolution is processed sequentially through the first splicing layer, the first segmentation layer, the first bottleneck layer, and the second bottleneck layer; the output of the second bottleneck layer, the output of the first splicing layer, and the output of the first segmentation layer are input together into the second splicing layer and then processed into the second convolutional layer to obtain the output of the first C3k2 module.

9. The dual-energy X-ray mineral detection method based on dual-backbone YOLOv11 according to claim 8, characterized in that: The third, fourth, sixth, seventh, and eleventh C3k2 modules have the same structure. The third C3k2 module includes a third splicing layer, a second segmentation layer, a first C3k module, a second C3k module, a fourth splicing layer, and a third convolutional layer. The processing flow of the third C3k2 module is as follows: the output of the first modulated deformable convolution is processed sequentially through the third splicing layer, the second segmentation layer, the first C3k module, and the second C3k module. The outputs of the second c3k module, the third splicing layer, and the second segmentation layer are input together into the fourth splicing layer and then processed in the third convolutional layer to obtain the output of the third C3k2 module.

10. The dual-energy X-ray mineral detection method based on dual-backbone YOLOv11 according to claim 9, characterized in that: The dual-backbone YOLOv11 model uses the SIoU loss function.

Images